To address the challenges of data silos and privacy in cross-institutional collaboration, this study introduces a secure data collaboration framework combining federated learning (FL) and differential privacy (DP).The framework enables collaborative model training by keeping data local while using client-side DP to counter privacy threats like membership inference attacks.An adaptive privacy budget allocation (APBA) strategy further optimises the utility-privacy balance.Evaluations on real educational datasets show the framework maintains strong privacy (attack success <5%), achieves a 95% F1 score -comparable to centralised trainingand improves communication efficiency by ~40%.This work provides a technical foundation for building secure and efficient cross-institutional platforms in innovation and entrepreneurship education.
Jiang et al. (Thu,) studied this question.